Radiation source individual identification method based on complex value contrast learning and short-time Fourier transform

Abstract

The invention relates to the technical field of radiation source identification, and discloses a radiation source individual identification method based on complex value comparative learning and short-time Fourier transform, which comprises the following steps: S1, carrying out short-time Fourier transform on I / Q channel signals, S2, carrying out data enhancement on a complex value spectrogram, and S3, carrying out data processing on the complex value spectrogram after data enhancement, s4, in the self-supervised pre-training stage, using a complex-valued convolutional neural network as an encoder for feature extraction; S5, entering a supervised fine-tuning training stage; S501, using marked data for training; and S6, using the marked data for a radiation source individual identification task. Through complex value comparison learning and deep feature extraction, frequency domain features of signals are effectively learned under an unsupervised condition, data enhancement and self-supervised pre-training improve the capability of distinguishing different types of signals, and for complex types of signals such as radar signals and communication signals, the model can accurately identify the features of the signals in different application scenarios, so that the accuracy of the model is improved. And identification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source identification, and specifically to a radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform. Background Art

[0002] The radiation source individual identification technology is one of the key technologies in modern wireless communication, electronic warfare, radar detection and other fields, aiming to identify radiation sources from different sources by analyzing the received signals of radio waves. With the increasing busyness and complexity of the radio spectrum, how to effectively extract key feature information from complex signals and then accurately identify and classify radiation sources has become an important research direction in current radio monitoring and electronic countermeasure systems.

[0003] Traditional radiation source identification methods mostly rely on feature extraction of frequency-domain or time-domain signals. These methods usually rely on manually designed features, such as signal modulation mode, frequency, amplitude, pulse characteristics, etc. However, with the increasing complexity of radiation source signals, traditional methods often face great challenges when dealing with diverse and highly complex signals. Especially when facing uncertainty factors such as signal noise and signal distortion, the robustness and adaptability of traditional feature extraction methods are poor. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform, which solves the problems that the performance of different types of signals varies greatly in different application scenarios, resulting in the model having no accuracy in different environments and conditions, and being unable to identify signal features due to the influence of noise, interference and signal distortion.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform, including the following steps: S1. Perform short-time Fourier transform on the I / Q channel signals. Generate a spectrogram by performing STFT on the I channel signal as the real part, and generate a spectrogram by performing STFT on the Q channel signal as the imaginary part. After combining the two, a set of complex-valued spectrograms is formed, and the frequency-domain characteristics of the signal are fully displayed through the spectrogram for subsequent feature extraction and signal analysis; S2. Perform data augmentation on the complex-valued spectrogram to generate positive and negative samples. By performing data augmentation processing on the complex-valued spectrogram, complex-valued image positive samples and negative samples are formed; S3. Input the data-augmented complex-valued spectrogram into a two-channel contrast learning model based on a complex-valued contrast learning framework for self-supervised pre-training; S4. Use a complex-valued convolutional neural network as an encoder feature extractor in the self-supervised pre-training stage and extract frequency-domain features; S5. After completing self-supervised pre-training, enter the supervised fine-tuning training stage. Load the best model weights from the self-supervised pre-training stage into the model, train using labeled data and the cross-entropy loss function, and lock the weights of the encoder part of the complex-valued CNN; S501. In the supervised fine-tuning training stage, do not perform data augmentation. Use labeled data for training, and update the parameters of the fully connected layer through backpropagation to improve the accuracy and generalization ability of the model in actual radiation source individual recognition; S6. After training, save the final model parameters and use them for the radiation source individual recognition task. The model can effectively identify different radiation source signals to reduce the dependence on labeled data.

[0006] Preferably, in S1, the spectrogram includes the amplitude spectrum and the phase spectrum, the complex-valued spectrogram includes the real part spectrogram and the imaginary part spectrogram, the signal frequency domain features include the center frequency, bandwidth, spectral peak, and spectral shape, the feature extraction includes time-frequency feature extraction, statistical feature extraction, local feature extraction, and transform-based feature extraction, and the signal analysis includes modulation mode recognition, target recognition, fault diagnosis, and signal quality assessment.

[0007] Preferably, in S2, the data augmentation includes geometric transformation and photometric transformation, the data augmentation process includes random parameter selection, combined transformation, and semantic consistency, the positive image samples include enhanced versions of the same complex-valued spectrogram and feature similarity, and the negative image samples include enhanced versions of different complex-valued spectrograms and feature differences.

[0008] Preferably, in S3, the contrastive learning model includes SimCLR, and the self-supervised pre-training includes encoding and projection, positive and negative sample contrast, and hyperparameter adjustment.

[0009] Preferably, in S4, the encoder feature extraction uses a complex-valued CNN, and the frequency domain features include basic spectral features, frequency statistical features, high-order frequency domain features, and phase-related features.

[0010] Preferably, in S5, the supervised fine-tuning training stage includes freezing and fine-tuning, loss calculation and optimization, and model evaluation and adjustment. The best model weights are used to measure the performance of the model in different training stages through various evaluation metrics, and the various evaluation metrics include the contrastive loss value and the accuracy on the validation set. The labeled data includes input features, labels, and data annotation methods.

[0011] Preferably, in S501, the parameters of the fully connected layer include the weight matrix and the bias term, and the radiation source individuals include different types of radiation sources and different individuals within the same type.

[0012] Preferably, in S6, the different radiation source signals include radar signals, communication signals, and signals characterized by signal features and application scenarios.

[0013] The present invention provides a method for identifying individual radiation sources based on complex-valued contrastive learning and short-time Fourier transform. It has the following beneficial effects: 1. Through complex-valued contrastive learning and deep feature extraction, the model can effectively learn the frequency-domain features of signals under unsupervised conditions, and improve the discrimination ability for different types of signals through data augmentation and self-supervised pre-training. For complex types of signals such as radar signals and communication signals, the model can accurately identify their features in different application scenarios, improving the recognition accuracy.

[0014] 2. By adopting data augmentation technology and a contrastive learning framework, the model's robustness can be enhanced on limited labeled data. Even in the face of noise or signal distortion, the model can maintain a high generalization ability and has the effect of adapting to various changes in different signals.

[0015] 3. Through the combination of self-supervised pre-training and a complex-valued convolutional neural network, the model first learns the features of signals with the help of large-scale unlabeled data, and then fine-tunes with a small amount of labeled data. This method reduces the dependence on a large amount of labeled data, enabling the model to still exhibit good recognition performance when there is less labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the method for identifying individual radiation sources based on complex-valued contrastive learning and short-time Fourier transform of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] Please refer to the attached Figure 1 , the embodiments of the present invention provide a method for identifying individual radiation sources based on complex-valued contrastive learning and short-time Fourier transform, including the following steps: S1. Perform short-time Fourier transform on the I / Q channel signals. Generate a spectrogram of the I channel signal through STFT as the real part, and generate a spectrogram of the Q channel signal through STFT as the imaginary part. After combining the two, a set of complex-valued spectrograms are formed to fully display the frequency-domain features of the signals for subsequent feature extraction and signal analysis; S2. Perform data augmentation on the complex-valued spectrogram to generate positive and negative samples. By performing data augmentation on the complex-valued spectrogram, positive and negative samples of complex-valued images are formed; S3. Input the data-augmented complex-valued spectrogram into a two-channel contrastive learning model based on a complex-valued contrastive learning framework for self-supervised pre-training; S4. Use a complex-valued convolutional neural network as an encoder for feature extraction and extract frequency-domain features during the self-supervised pre-training stage; S5. After completing the self-supervised pre-training, enter the supervised fine-tuning training stage. Load the best model weights from the self-supervised pre-training stage into the model, train using labeled data and the cross-entropy loss function, and lock the weights of the encoder part of the complex-valued CNN; S501. During the supervised fine-tuning training stage, do not perform data augmentation. Use labeled data for training and update the parameters of the fully connected layer through backpropagation to improve the accuracy and generalization ability of the model in actual radiation source individual recognition; S6. After training, save the final model parameters and use them for the radiation source individual recognition task. The model can effectively identify different radiation source signals, reducing the dependence on labeled data.

[0019] In S1, the spectrogram includes the amplitude spectrum and the phase spectrum. The complex-valued spectrogram includes the real-part spectrogram and the imaginary-part spectrogram. The signal frequency-domain features include the center frequency, bandwidth, spectral peak, and spectral shape. Feature extraction includes time-frequency feature extraction, statistical feature extraction, local feature extraction, and transform-based feature extraction. Signal analysis includes modulation mode recognition, target recognition, fault diagnosis, and signal quality assessment.

[0020] Specifically, the amplitude spectrum shows the energy distribution of the signal spectrum, and the phase spectrum provides the phase information in the frequency domain. The STFT spectrograms of the I-channel and Q-channel signals are used as the real part and the imaginary part of the complex-valued spectrogram respectively to form a complex frequency-domain representation, revealing features such as the center frequency, bandwidth, spectral peak, and spectral shape of the signal, which are helpful for identifying different radiation sources; The feature extraction methods include time-frequency feature extraction, statistical feature extraction, local feature extraction, and transform-based feature extraction. Time-frequency feature extraction captures the time-frequency changes of the signal through STFT; statistical feature extraction describes the global characteristics of the signal through mean, variance, etc.; local feature extraction focuses on important signal patterns in the spectrum; transform-based feature extraction uses wavelet transform, etc. to further explore the signal characteristics. These methods work together to comprehensively improve the recognition accuracy; In the signal analysis stage, in combination with tasks such as modulation mode recognition, target recognition, fault diagnosis, and signal quality assessment, in-depth analysis of signals is carried out. Modulation mode recognition is used to distinguish different modulation types, target recognition locates the signal source by matching with a known target library, fault diagnosis helps to detect equipment or system faults, and signal quality assessment judges the stability of signal transmission. Through complex-valued contrastive learning and in-depth analysis of spectrograms, the dependence on labeled data is reduced, and the accuracy and generalization ability of signal recognition are improved.

[0021] In S2, data augmentation includes geometric transformation and photometric transformation. The data augmentation process includes random parameter selection, combined transformation, and semantic consistency. Image positive samples include enhanced versions of the same complex-valued spectrogram and feature similarity, and image negative samples include enhanced versions of different complex-valued spectrograms and feature differences.

[0022] Specifically, data augmentation includes geometric transformation and photometric transformation, aiming to generate new samples by transforming complex-valued spectrograms in different ways to enrich the training set. Geometric transformation includes operations such as rotation, scaling, and translation, which can simulate the changes of signals in different directions and scales; photometric transformation includes adjustments such as brightness, contrast, and noise, enhancing the model's adaptability to signal intensity and noise changes. The data augmentation process generates enhanced samples with semantic consistency by randomly selecting transformation parameters and combining different transformation methods. Image positive samples are enhanced versions generated by transforming the same complex-valued spectrogram, and these enhanced versions have high similarity in features and can reflect the robustness of signals under different transformations. Image negative samples are enhanced versions generated by transforming different complex-valued spectrograms and have large feature differences, which are used to help the model distinguish different types of signals. Data augmentation not only expands the number of training samples but also improves the model's adaptability to different signal deformations, thus enhancing the accuracy and robustness of individual emitter recognition.

[0023] In S3, the contrastive learning model includes SimCLR, and self-supervised pre-training includes encoding and projection, positive and negative sample contrast, and hyperparameter adjustment.

[0024] Specifically, the feature representation of data is learned in an unsupervised manner. In the process of self-supervised pre-training, first, the input complex-valued spectrogram is converted into a low-dimensional feature space through encoding and projection. The encoder is responsible for extracting the latent features in the spectrogram, and the projection head maps these features into a space where the distance between similar samples is close and the distance between different samples is far. The positive and negative sample contrast is the core mechanism in SimCLR. In self-supervised learning, positive sample pairs refer to different augmented versions of the same complex-valued spectrogram, while negative sample pairs come from different complex-valued spectrograms. By maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, the model can learn effective feature representations in the data. This contrastive learning framework can force the model to capture the discriminative features between different signals; During the pre-training process, hyperparameters also need to be adjusted to optimize the training effect of the contrastive learning model. Adjustments of hyperparameters such as the learning rate, batch size, and dimension of the projection head can effectively improve the quality of feature learning. By reasonably selecting hyperparameters, the model can efficiently learn deep-level feature representations of signals under unsupervised conditions. The SimCLR model can learn efficient feature representations in complex-valued spectrograms during the self-supervised pre-training stage, providing strong feature support for subsequent supervised fine-tuning and radiation source individual identification tasks.

[0025] In S4, for encoder feature extraction, complex-valued CNN is used. The frequency domain features include basic spectral features, frequency statistical features, high-order frequency domain features, and phase-related features.

[0026] Specifically, complex-valued CNN is different from traditional real-valued CNN. Its convolution operation can process both the real and imaginary parts of complex numbers simultaneously, thus more accurately capturing the frequency domain characteristics of signals. The extraction of frequency domain features includes basic spectral features, frequency statistical features, high-order frequency domain features, and phase-related features. Basic spectral features include center frequency, bandwidth, spectral peak, etc., which can provide the model with global frequency domain information of signals and help distinguish different signal types. Frequency statistical features further quantify the frequency distribution characteristics of signals by calculating statistics such as the mean, variance, skewness, and kurtosis of the signal spectrum, which is very important for identifying the detailed features of different radiation source signals. High-order frequency domain features extract potential complex patterns in signals by analyzing higher-order information of the spectrum, such as higher-order statistics of the spectrum and non-linear changes of the spectrum, which helps to process non-linear and complex signals. Phase-related features reveal the phase relationship between signals by analyzing the phase spectrum of the signal, which plays an important role in analyzing modulation methods and other signal characteristics.

[0027] Complex-valued CNN gradually extracts these frequency domain features through convolutional layers and maps them into deep-level feature representations, further enhancing the model's identification ability. Through complex-valued CNN, the model can effectively capture multi-dimensional information in the signal spectrum, thus providing accurate frequency domain feature support for subsequent radiation source individual identification tasks.

[0028] In S5, the supervised fine-tuning training stage includes freezing and fine-tuning, loss calculation and optimization, and model evaluation and adjustment. The best model weights are used to measure the performance of the model at different training stages through various evaluation metrics, including contrast loss values and accuracy on the validation set. The labeled data includes input features and labels and the data annotation method.

[0029] Specifically, some weights of the complex-valued convolutional neural network (CNN) encoder are frozen to retain the effective features learned in the self-supervised pre-training stage, while the parameters of the fully connected layer are fine-tuned to further improve the performance of the model on specific tasks. The combination of freezing and fine-tuning can ensure that the encoder part is not disturbed during the model training process, retaining the useful feature information learned from the unlabeled data, while the fine-tuning of the fully connected layer can refine the model through the labeled data and enhance its recognition ability in actual tasks. Loss calculation and optimization are the key in the supervised fine-tuning stage. Usually, the cross-entropy loss function is used to calculate the difference between the model output and the true label, and then the network weights are updated through the backpropagation algorithm. Optimization algorithms such as Adam or SGD can be used to minimize the loss value, thereby improving the accuracy and robustness of the model. During the model evaluation and adjustment process, the best model weights are used to measure the performance of the model at different training stages. This process is carried out through a variety of evaluation metrics, mainly including contrast loss values and accuracy on the validation set. The contrast loss value reflects the ability of the model to distinguish between positive and negative samples during the learning process, while the accuracy on the validation set measures the generalization ability of the model on unseen data. Through these evaluation metrics, the performance of the model during training can be effectively judged, and necessary adjustments can be made, such as adjusting the learning rate, changing the optimization strategy, etc., to further improve the model performance. The role of the labeled data at this stage is to provide true labels for the model for supervised learning. The labeled data includes input features and labels. The input features are complex-valued spectrograms and the frequency-domain features extracted from them, while the labels are the corresponding radiation source types or other recognition targets. The data annotation method can adopt manual annotation, automated annotation or semi-automated annotation, etc., to ensure the accuracy and reliability of the labeled data. The supervised fine-tuning training stage can enable the model to make full use of the labeled data in actual tasks, improve the recognition accuracy, and further optimize the model through the feedback of the evaluation metrics, and finally obtain a high-performance radiation source individual recognition model.

[0030] In S501, the parameters of the fully connected layer include the weight matrix and the bias term. The radiation source individuals include different types of radiation sources and different individuals within the same type.

[0031] Specifically, the fully connected layer is responsible for converting the frequency-domain features extracted by the complex-valued convolutional neural network (CNN) into the final output categories. The weight matrix and bias terms are optimized through the backpropagation algorithm, enabling the model to output correct class predictions based on the input features. The weight matrix is used to control the mapping relationship between the input features and the output categories, while the bias terms are used to adjust the prediction results of the model. By minimizing the loss function (such as cross-entropy loss) and updating these parameters, the model can gradually improve the accuracy of identifying radiation sources; The definition of individual radiation sources involves two aspects: Firstly, individual radiation sources include different types of radiation sources, signal sources with different frequencies, modulation methods, or other characteristics; Secondly, different individuals within the same type are also regarded as different individual radiation sources. For example, there may be multiple different devices or signal sources of the same type of radiation source, and the frequency-domain characteristics of each signal source vary to some extent. Therefore, the model needs to be able to distinguish different individuals within the same type. By fine-tuning the parameters of the fully connected layer, when the model performs individual radiation source identification, it can more precisely capture the characteristic differences between different types and different individuals, thereby improving the accuracy and reliability in practical applications.

[0032] In S6, different radiation source signals include radar signals, communication signals, and signals characterized by signal features and application scenario features.

[0033] Specifically, effective discrimination is carried out according to the characteristics of the signal and the different application scenarios. Radar signals and communication signals have their own unique frequency-domain characteristics and time-domain characteristics. Radar signals usually have relatively special pulse characteristics, while communication signals may contain complex modulation methods and signal forms. In order for the model to accurately identify different radiation source signals, the model not only needs to analyze the frequency-domain characteristics of the signal, such as the center frequency, bandwidth, spectral peak, etc., but also consider the application scenario characteristics. In radar applications, characteristics such as the time-frequency distribution and pulse repetition frequency (PRF) of the signal are crucial for identifying radar signals; while in the identification of communication signals, modulation methods, symbol rate, and the error tolerance of the signal are important discriminant bases. The model in the present invention can perform intelligent classification according to the input signal characteristics (such as spectral characteristics, modulation methods, etc.) and application scenario characteristics (such as radar scenarios, communication environments, etc.), accurately identify and distinguish different radiation source signals, which enables the model to process a variety of different types of signals in practical tasks and maintain high recognition accuracy and generalization ability in complex environments.

[0034] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform, characterized in that: The following steps are involved: S1. Perform short-time Fourier transform on the I / Q channel signals, use the spectrum generated by STFT of the I channel signal as the real part, and use the spectrum generated by STFT of the Q channel signal as the imaginary part. The two are combined to form a set of complex-valued spectrum graphs, which fully display the frequency domain characteristics of the signal for subsequent feature extraction and signal analysis; S2. Perform data enhancement on the complex-valued spectrum graph to generate positive and negative samples. Perform data enhancement on the complex-valued spectrum graph to form positive and negative samples of complex-valued images. S3, inputting the data-enhanced complex-valued spectrogram into a dual-channel contrastive learning model based on a complex-valued contrastive learning framework for self-supervised pre-training; S4, using complex-valued convolutional neural network as encoder feature extraction in the self-supervised pre-training stage, and extracting frequency domain features; S5. After completing the self-supervised pre-training, enter the supervised fine-tuning training phase, load the best model weights in the self-supervised pre-training phase into the model, train it using labeled data and the cross entropy loss function, and lock the encoder weights of the complex-valued CNN; S501. In the supervised fine-tuning training stage, no data enhancement is performed, labeled data is used for training, and the parameters of the fully connected layer are updated through back propagation to improve the accuracy and generalization ability of the model in the actual individual identification of radiation sources; S6. After training, the final model parameters are saved and used for the task of individual radiation source identification. The model can effectively identify signals of different radiation sources and reduce the dependence on labeled data.

2. The radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform according to claim 1 is characterized in that: In S1, the spectrum diagram includes an amplitude spectrum and a phase spectrum, the complex-valued spectrum diagram includes a real spectrum diagram and an imaginary spectrum diagram, the signal frequency domain features include a center frequency, a bandwidth, a spectrum peak and a spectrum shape, the feature extraction includes time-frequency feature extraction, statistical feature extraction, local feature extraction and transformation-based feature extraction, and the signal analysis includes modulation mode recognition, target recognition, fault diagnosis and signal quality assessment.

3. The radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform according to claim 1 is characterized in that: In S2, the data enhancement includes geometric transformation and photometric transformation, the data enhancement processing includes random parameter selection, combined transformation and semantic consistency, the image positive sample includes an enhanced version of the same complex-valued spectrum graph and feature similarity, and the image negative sample includes an enhanced version of different complex-valued spectrum graphs and feature difference.

4. The radiation source individual identification method based on complex-valued contrast learning and short-time Fourier transform according to claim 1 is characterized in that: In S3, the contrastive learning model includes SimCLR, and the self-supervised pre-training includes encoding and projection, positive and negative sample comparison, and hyperparameter adjustment.

5. The method for identifying radiation sources based on complex-valued contrastive learning and short-time Fourier transform according to claim 1, characterized in that: In S4, the encoder feature extraction adopts complex-valued CNN, and the frequency domain features include basic spectrum features, frequency statistical features, high-order frequency domain features and phase-related features.

6. The method for identifying radiation sources based on complex-valued contrastive learning and short-time Fourier transform according to claim 1, characterized in that: In S5, the supervised fine-tuning training stage includes freezing and fine-tuning, loss calculation and optimization, and model evaluation and adjustment. The optimal model weight is used to measure the performance of the model at different training stages through various evaluation indicators. The various evaluation indicators include comparative loss value and accuracy on the validation set. The labeled data includes input features and labels and data annotation methods.

7. The method for identifying radiation sources based on complex-valued contrastive learning and short-time Fourier transform according to claim 1, characterized in that: In S501, the parameters of the fully connected layer include a weight matrix and a bias term, and the radiation source individuals include radiation sources of different types and different individuals of the same type.

8. The method for identifying radiation sources based on complex-valued contrastive learning and short-time Fourier transform according to claim 1, characterized in that: In S6, the different radiation source signals include radar signals, communication signals, and signals based on signal characteristics and application scenario characteristics.

Images